An Improved Cloud Image-Solar Irradiance Ultra-Short-Term Forecasting Model Based on DRL and RBFNN

The motion characteristics of clouds can lead to sudden changes in photovoltaic power. And to accurately predict the power of distributed photovoltaics, it is first necessary to accurately predict the ultra-short-term solar irradiance received at the surface with construction of an ultra-short-term solar irradiance forecasting model based on cloud image features. During the cloud image data processing, the data is firstly standardized, and the intra-day variability of it is reduced by suppressing the influence of the solar altitude angle on the pixel intensity of cloud images. On this basis, a centroid algorithm is used to predict the cloud images for a ultra-short period in the future; then Deep Residual Learning (DRL) is proposed to extract cloud image features. These features are combined with other influencing factors to form an input feature vector, and put into Radial Basis Function Neural Network (RBFNN) model to obtain the surface radiation forecasting results. Experimental results show that the model can accurately predict the intensity of solar radiation received by the surface of a set area.